Re: [PATCH v2] xfs: use a dedicated SLAB cache for sector sized buffer data

[Brian Foster <bfoster@redhat.com>]

On Thu, Dec 06, 2018 at 12:26:14PM -0800, Darrick J. Wong wrote:
> On Thu, Dec 06, 2018 at 09:11:06PM +0100, Christoph Hellwig wrote:
> > On Thu, Dec 06, 2018 at 10:11:39AM -0800, Darrick J. Wong wrote:
> > > I think this has the same problem that Dave complained about back in
> > > October -- if I have a hard disk with 512b sectors, a xfs with 4k
> > > blocks, and mount it on a machine with 64k pages, then single-sector
> > > buffers will get a small slab allocation, but single-fsb buffers will
> > > now fall back to grabbing a 64k page to hold 4k worth of data.
> > > 
> > > Even if we add a second slab for single-fsb blocks, we'll run into the
> > > same problem if the filesystem contains multi-block directory blocks, so
> > > then we'd potentially need a third slab, and... is there a way to ask a
> > > slab allocator for multiple contiguous objects totalling less than
> > > PAGE_SIZE bytes?
> > 
> > No, unfortunately there is not.  And the sad part is that generally it
> > will give you aligned ones, just with slub debugging turned on it won't.
> 
> Hmmm.  I guess we could declare three slabs then -- sector size, fsb size,
> and dir fsb size?  This is getting kinda spidery now though.  Is three
> too weird?
> 

You can add inode cluster size into the mix as well (see the snipped off
portion of my previous reply that I guess was missed). ;P

Brian

> --D